To address the problem of inadequate background motion compensation in complex scenarios for infrared images, this paper introduces a novel image registration algorithm using the LGB (Location-Gray-BEBLID) descriptor. Initially, a quadtree algorithm is applied to remove excessive feature points, efficiently managing the issue of overcrowded feature points. The proposed LGB descriptor combines feature point location and grayscale data, significantly improving the precision of feature point matching. Next, a block matching technique is utilized for initial feature point matching, followed by a refinement process to enhance the matching outcomes. Ultimately, the background motion is calculated using the matched feature points, enabling effective compensation. Experimental findings demonstrate that this algorithm surpasses existing methods in compensation performance on various test image sequences, with exceptional real-time capabilities, reducing processing time by around 59% compared to the traditional ORB algorithm. This advancement in background motion compensation is a significant step towards enhancing infrared dim target detection.Background motion compensation is a crucial technique in the processing of infrared images with complex backgrounds, offering significant practical value. In certain intricate infrared scenes, targets are small, have low energy, and are often heavily obscured by background clutter. Traditional spatial domain infrared small target detection algorithms, such as the Top-Hat algorithm and the MPCM (Multiscale Patch-based Contrast Measure) algorithm, tend to perform inadequately in these challenging conditions. Background motion compensation effectively mitigates these algorithms’ shortcomings when handling complex background images. This compensation is usually achieved by calculating the parameters of the background motion. By differencing the current frame image with the previous frame image after applying background motion compensation, the interference from the background can be significantly reduced, thereby improving the conditions for detecting infrared dim targets. Among various background motion compensation techniques, methods based on image registration are widely used due to their high accuracy.The core steps of image registration include feature point detection and feature point descriptor generation. This paper introduces an innovative image registration method based on LGB descriptors for effective background motion compensation (Fig.1). To address the issue of redundant feature points, a quadtree algorithm is utilized (Fig.3). The LGB (Location-Gray-BEBLID) descriptor is introduced for the first time, enhancing the BEBLID descriptor by incorporating both the position and grayscale information of feature points, thereby significantly improving the precision and discriminability of feature point matching (Fig.4). During the feature point matching phase, a block matching strategy is implemented. This strategy constructs a hash function to map the coordinates of feature points in different regions to distinct key values, limiting the matching process to feature points with the same key values. This approach effectively reduces the search range and time required for matching (Fig.8). The Random Sample Consensus (RANSAC) algorithm is then employed to refine and optimize the matching results. By calculating background motion parameters from the feature point coordinate information, the algorithm successfully achieves background motion compensation.Fig. 13 analyzes the efficiency of different matching strategies. Points near the upper left indicate shorter processing times and higher correct matching rates, indicating better performance. Dividing the image into 2×2 blocks strikes the best balance between performance and efficiency. Tab. 4 compares OLE, NFAM, and correct matching rates of various algorithms, with optimal results in bold. The proposed algorithm shows a higher average correct matching rate, indicating richer information from correct feature point pairs in Equation (9), leading to more accurate parameters and better background motion compensation. The Ours* algorithm outperforms others in OLE, NFAM, and correct matching rate, suggesting the proposed descriptor is more effective in capturing feature point characteristics, resulting in more precise matching and superior background motion compensation compared to other descriptors. The comparison between the ORB and BRIEF algorithms demonstrates the effectiveness of the quadtree algorithm.Tab.5 details time consumption, where the proposed quadtree algorithm significantly reduces feature detection time. The proposed descriptor generation is the quickest, 38% faster than BEBLID, due to its efficient location and grayscale computation. The proposed block matching reduces matching time by 58% versus point-by-point methods. The proposed algorithm is the most time-efficient, reducing total time by 43% compared to BEBLID and 59% compared to ORB.To address the challenge of inadequate background motion compensation in complex infrared scenes, this paper introduces a novel image registration method based on LGB descriptors. Initially, the algorithm employs a quadtree technique to optimize the distribution of feature points, mitigating the issue of redundancy. Building upon the BEBLID descriptor, the LGB descriptor is proposed, incorporating both the position and grayscale information of feature points. This enhancement significantly improves the precision and discriminability of feature point matching. Then the matching time is reduced by the proposed block matching strategy. The RANSAC algorithm is then utilized to refine and enhance the matching results. By calculating background motion parameters from the feature point information, effective background motion compensation is achieved, thereby eliminating background interference. Experimental results demonstrate that this algorithm excels in background motion compensation for complex infrared images, outperforming other methods in terms of the OLE, NFAM and correct matching rate metrics. The time efficiency of generating descriptors is impressive, showing a reduction of approximately 38% compared to the BEBLID algorithm. Moreover, the proposed algorithm reduces matching time by approximately 58% compared to point-by-point matching.The total processing time of our algorithm is about 59% less than that of the ORB algorithm. This method not only improves the effectiveness of background motion compensation in complex infrared scenes but also enhances its efficiency, paving the way for more effective infrared dim target detection.